<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Shoeb Reza | Arwahyn AI</title>
    <description>The latest articles on DEV Community by Shoeb Reza | Arwahyn AI (@arwahyn_ai_a9134e083fcc3d).</description>
    <link>https://dev.to/arwahyn_ai_a9134e083fcc3d</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4254432%2F895b51bb-9307-4444-b132-52ca20dd27e9.png</url>
      <title>DEV Community: Shoeb Reza | Arwahyn AI</title>
      <link>https://dev.to/arwahyn_ai_a9134e083fcc3d</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arwahyn_ai_a9134e083fcc3d"/>
    <language>en</language>
    <item>
      <title>We built an AI news desk that writes in Bangla twice a day, clusters duplicate stories, and credits every source</title>
      <dc:creator>Shoeb Reza | Arwahyn AI</dc:creator>
      <pubDate>Thu, 19 Nov 2026 20:20:42 +0000</pubDate>
      <link>https://dev.to/arwahyn_ai_a9134e083fcc3d/we-built-an-ai-news-desk-that-writes-in-bangla-twice-a-day-clusters-duplicate-stories-and-credits-5747</link>
      <guid>https://dev.to/arwahyn_ai_a9134e083fcc3d/we-built-an-ai-news-desk-that-writes-in-bangla-twice-a-day-clusters-duplicate-stories-and-credits-5747</guid>
      <description>&lt;p&gt;Most AI news in Bangla is either a translated headline with no context or a copied article. We wanted something closer to a morning newsletter: what happened, why it matters, and what it changes for someone working in Bangladesh, written in Bangla that reads like a person wrote it.&lt;/p&gt;

&lt;p&gt;So our academy runs its own small news desk:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;A collector&lt;/strong&gt; checks a list of tech and AI news sources every 30 minutes and scores each headline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two editorial runs a day&lt;/strong&gt;, at 8 AM and 7 PM Dhaka time, the way a newspaper has a morning and an evening edition. Each run writes at most three stories.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The bar rises for each extra story.&lt;/strong&gt; The first story only needs to clear a basic score. The second and third have to be genuinely big: well above a typical week's scores, or covered by two or more outlets.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Duplicate stories are clustered.&lt;/strong&gt; When five outlets cover the same launch, the desk writes it once and credits all five.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anything covered in the last 72 hours is skipped&lt;/strong&gt;, so the same story doesn't come back with a new headline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every article credits its sources by name and link.&lt;/strong&gt; Thumbnails are made fresh, not copied from the source.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A fixer turns Bangla-script tech names back into English letters&lt;/strong&gt;, because "ওপেনএআই" in a headline helps nobody find anything.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;On a slow news day it stops early and logs "slow news, stopped at one story" rather than padding. That's the part I'm happiest with, because a desk that always publishes three stories ends up publishing three weak ones.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stack:&lt;/strong&gt; Supabase (Postgres, edge functions, cron), Gemini for writing and images, Astro on Vercel for the site.&lt;/p&gt;

&lt;p&gt;You can read the result at &lt;a href="https://academy.arwahynai.com/news" rel="noopener noreferrer"&gt;academy.arwahynai.com/news&lt;/a&gt;.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;I run &lt;a href="https://arwahynai.com" rel="noopener noreferrer"&gt;Arwahyn AI&lt;/a&gt;, an AI automation agency in Dhaka.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>supabase</category>
      <category>sideprojects</category>
      <category>automation</category>
    </item>
    <item>
      <title>A private ChatGPT for the office with Ollama + Open WebUI, where nothing leaves your machine</title>
      <dc:creator>Shoeb Reza | Arwahyn AI</dc:creator>
      <pubDate>Thu, 19 Nov 2026 20:20:22 +0000</pubDate>
      <link>https://dev.to/arwahyn_ai_a9134e083fcc3d/a-private-chatgpt-for-the-office-with-ollama-open-webui-where-nothing-leaves-your-machine-1668</link>
      <guid>https://dev.to/arwahyn_ai_a9134e083fcc3d/a-private-chatgpt-for-the-office-with-ollama-open-webui-where-nothing-leaves-your-machine-1668</guid>
      <description>&lt;p&gt;A question we keep getting from business owners: "my staff are already pasting customer emails and contracts into ChatGPT, can we stop that without banning AI?" For a lot of offices the honest answer is a local model on one decent machine, so nothing leaves the building.&lt;/p&gt;

&lt;p&gt;It's like having a typist in the back room instead of posting your documents to a typing service across town. They're a bit slower, but the papers never leave the office.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you need
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;One machine with a recent GPU (or an Apple Silicon Mac with enough memory). A small model also runs on CPU for light use, just slowly.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ollama&lt;/strong&gt; to run the models&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open WebUI&lt;/strong&gt; for the chat screen your team uses in a browser&lt;/li&gt;
&lt;li&gt;Docker for Open WebUI&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Setup
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Install Ollama from its site and pull a model sized to your hardware. Start small and go up only if the answers aren't good enough.&lt;/li&gt;
&lt;li&gt;Run Open WebUI in Docker and point it at Ollama.&lt;/li&gt;
&lt;li&gt;Create the admin account, then turn off open sign-ups so only people you invite can use it.&lt;/li&gt;
&lt;li&gt;Give the team the machine's local address. It works over the office network, and nothing goes out to the internet unless you add tools that do.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Where it works well
&lt;/h2&gt;

&lt;p&gt;Summarising long emails, rewriting replies politely, pulling details out of documents and drafting first versions. Most office writing doesn't need the biggest model in the world.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it doesn't
&lt;/h2&gt;

&lt;p&gt;Hard reasoning, very long documents on small hardware, and anything that needs fresh knowledge from the internet. For those, use a paid AI service under a business plan whose data terms you've actually read.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rules we give clients
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;One shared machine, set up properly, beats ten laptops each running something different.&lt;/li&gt;
&lt;li&gt;Back up the Open WebUI data folder, because chats and settings live there.&lt;/li&gt;
&lt;li&gt;Write down which model is approved for which kind of document.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;em&gt;I run &lt;a href="https://arwahynai.com" rel="noopener noreferrer"&gt;Arwahyn AI&lt;/a&gt;, an AI automation agency in Dhaka. Bangla setup guides for &lt;a href="https://arwahynai.com/bn/blog/ollama-local-ai-models-free-setup-bangla" rel="noopener noreferrer"&gt;Ollama&lt;/a&gt; and &lt;a href="https://arwahynai.com/bn/blog/open-webui-own-chatgpt-setup-bangla" rel="noopener noreferrer"&gt;Open WebUI&lt;/a&gt; are on our blog.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>ollama</category>
      <category>privacy</category>
      <category>selfhosted</category>
    </item>
    <item>
      <title>I describe an n8n workflow in plain English and Claude Code builds it: the setup and the 4 rules that stop it breaking things</title>
      <dc:creator>Shoeb Reza | Arwahyn AI</dc:creator>
      <pubDate>Thu, 19 Nov 2026 20:01:31 +0000</pubDate>
      <link>https://dev.to/arwahyn_ai_a9134e083fcc3d/i-describe-an-n8n-workflow-in-plain-english-and-claude-code-builds-it-the-setup-and-the-4-rules-1epe</link>
      <guid>https://dev.to/arwahyn_ai_a9134e083fcc3d/i-describe-an-n8n-workflow-in-plain-english-and-claude-code-builds-it-the-setup-and-the-4-rules-1epe</guid>
      <description>&lt;p&gt;We build n8n automations for clients every week, and the slowest part used to be clicking nodes together and hunting for the right field names. Now I describe the workflow to Claude Code, it builds it in our n8n instance, and I review it, a bit like dictating a letter to someone who types faster than you and then reading it before it goes out.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you need
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Claude Code installed&lt;/li&gt;
&lt;li&gt;An n8n instance you can reach (self-hosted or cloud) and an API key from &lt;strong&gt;n8n → Settings → n8n API&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;The open-source &lt;strong&gt;n8n-MCP&lt;/strong&gt; server, which gives Claude the docs and schemas for every n8n node, plus tools to create and update workflows&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Setup
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Add the n8n-MCP server to Claude Code with your n8n URL and API key as environment variables. The exact command is in the n8n-MCP README; it's one &lt;code&gt;claude mcp add …&lt;/code&gt; line.&lt;/li&gt;
&lt;li&gt;Restart Claude Code and ask it to list your workflows. If that works, it's connected.&lt;/li&gt;
&lt;li&gt;Keep the API key in an environment variable, never in a prompt or in a file you commit.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The 4 rules we give it
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Search the node docs before using a node.&lt;/strong&gt; Without this it guesses parameter names from memory and gets them half right.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate before saving.&lt;/strong&gt; n8n-MCP has a validate tool, and the workflow doesn't get created until validation is clean.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Never touch a live workflow directly.&lt;/strong&gt; It builds a copy, we test it with pinned sample data, and then we swap.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No real customer messages while testing.&lt;/strong&gt; Test data only. This one matters most for anything that sends WhatsApp messages or email.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  A prompt that works well
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;Build an n8n workflow: when a new row is added to the "Leads" Google Sheet, wait 5 minutes, send a WhatsApp message to the number in column C using the template in column D, and log the result in column E. Search the node docs first, validate before saving, and don't activate it.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Where it still needs a human
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Credentials:&lt;/strong&gt; you connect them in n8n yourself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anything with money, prices or customer promises:&lt;/strong&gt; read the prompt the AI node uses, line by line.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Error handling:&lt;/strong&gt; ask for it explicitly, or you'll get a happy-path workflow that fails silently.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If people want, I can share the system prompt we keep in our CLAUDE.md for n8n work.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;I run &lt;a href="https://arwahynai.com" rel="noopener noreferrer"&gt;Arwahyn AI&lt;/a&gt;, an AI automation agency in Dhaka. There's a step-by-step Bangla version of this setup &lt;a href="https://arwahynai.com/bn/blog/n8n-mcp-claude-code-builds-n8n-workflows-setup-bangla" rel="noopener noreferrer"&gt;on our blog&lt;/a&gt;, and we run free live AI automation classes in Bangla every Friday and Saturday at 9:30 PM Dhaka time at &lt;a href="https://academy.arwahynai.com" rel="noopener noreferrer"&gt;Arwahyn Academy&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>n8n</category>
      <category>ai</category>
      <category>automation</category>
      <category>claude</category>
    </item>
  </channel>
</rss>
